Applying a self-organizing map to sensor-array characterization

As a basic application of neural networks, the authors implemented a self-organizing map (SOM) as an algorithm to classify the response vectors from a sensor array exposed to various chemical vapors. Our chemical sensing system consists of an array of piezoelectric quartz-crystal microbalance (QCM) sensors, each coated with a different polymer membrane. Typically, statistical analysis are employed to characterize the sensor response to various gases and to classify each individual gas. However, because the sorption-desorption cycle can require a long time to come to equilibrium, the initial vectors do not contain much unique information. We replaced principal-component analysis with the self-organizing map as a visual method of finding the time at which the sensor-array signals become unique and of estimating the quality of the extracted features. In addition, we found that the SOM can accurately classify response vectors faster than principal-component analysis.